Paper Overview
Research Area: NLP Authors: Wei Zhou, Xiongwei Zhu, Zelin Xu Published: 2026-06-27 arXiv: 2606.27377
Abstract
Modern image generation demands a single model that unifies diverse capabilities, including text-to-image (T2I), local editing, and global editing. However, these capabilities are rarely naturally aligned and often conflict. For instance, editing tends to degrade T2I performance, while global and local editing interfere with each other. Consequently, effectively composing these capabilities has become a central challenge for image generation model training. To tackle this, the authors introduce DanceOPD, an on-policy generative field distillation framework that ensures each capability learns from the model's own generated distribution, avoiding misalignment and conflicts between capabilities.
Key Points
- Proposes a unified model covering T2I generation, local editing, and global editing
- Identifies the core problem: these capabilities conflict during joint training (e.g., editing degrades T2I quality)
- Uses an on-policy generative field distillation approach so each capability learns from the model's own generated distribution
- Paper link: https://arxiv.org/abs/2606.27377